Wildlife recognition from camera trap data using computer vision algorithms

Abstract

Camera trap, a digital camera that is automatically triggered by activities around, has been widely used in wildlife conservation for decades to capture animals on film for later analysis. With increasingly available vision data (photos and videos) from camera traps in recent years, it becomes prohibitively costly to manually extract useful information from these data. In this project, I aim to help automate the process of knowledge extraction from the camera trap data with the help of deep learning models. Specifically, a popular convolutional neural network (CNN) architecture called YOLOv3 was used as the pre-trained model through transfer learning. The model was then fine-tuned on thousands of camera trap images that I primitively obtained from a crowdsourced Zooniverse dataset and subsequently labeled using an object tagging tool. Compared to previously proposed work of wildlife recognition, my model further performs wildlife detection by locating the object detected and adding a bounding box in addition to identifying the species. As a result, the trained model is applied to photos of wildlife taken by myself for predictions, and results show that the model is able to accurately and confidently classify and locate multiple wildlife in both photos and real-time videos.